Fix init_get_thread_local()

Fix init_get_thread_local()

Previously, init_get_thread_local() pushed the thread event handler
list onto the global register before calling CRYPTO_THREAD_set_local(),
and when the latter failed, forgot to pop the list from the stack again.

Instead of cleaning the stack on error, this commit avoids the situation
entirely by postponing the push operation until all other operations
succeeded. This reordering also significantly reduces the scope of the
critical section.

Another simplification of the code is achieved by moving the push operation
onto the register (which is disabled in FIPS mode) into a separate function.

Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9401)